We're seeing a roughly 20% increase in memory utilization with the 5.4 RCs compared to 5.3 which we can not explain. Here's what we found so far.

From logs we see this happening during extraction, parsing.

It happens not only on bike but also on car (bike profile does not include lanes or turn restrictions but (!) the restriction parser is still getting called always returning an empty optional).

It seems to happen on larger datasets only, e.g. on the planet. But not on smaller ones.

#2704 => String deduplication map size looks reasonable; dumping some stats about the map did not reveal anything (e.g. a lot of hash collisions).

Extractor allbacks are explicitly getting allocated and destroyed again.

Extraction containers are scoped to the extractor and are cleaned up once we serialized the data out.

Other possibilities we need to investigate:

Touching the restriction parser code: #2829

Changing the scripting environment: 1309dd2 - currently running a staging stack changeset reverted. Restriction parser only need the scripting environment in the ctor for getting turn restrictions from profiles.

Attach heap profiler (e.g. GPerfTools) and run on planet: we only see this behavior on larger datasets.